MOTIVATION LETTER Drug resistance is the single largest cause of treatment failure in oncology and infectious disease. Predicting which mutations will confer resistance before they emerge in a patient would allow clinicians to choose a second-line therapy proactively and would let drug developers design molecules that evade resistance from the start. Current computational tools require a protein crystal structure to make a prediction, which means they cover at most 18 percent of clinically relevant mutations. The remaining 82 percent are invisible to those methods. My venture solves that gap. I am a pharmacist who retrained as a machine-learning engineer, and I have built a prediction pipeline that uses ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints and a Random Forest classifier. No crystal structure is needed. On the Platinum benchmark of 553 mutations, evaluated with protein-grouped cross-validation, the model achieves an AUROC of 0.804 plus or minus 0.025. That beats the published state of the art, mCSM-lig, which scores approximately 0.70. My model covers 100 percent of mutations in the test set, not 18 percent. Turn.io 2026 Health and AI Accelerator is the right programme for this stage because it explicitly seeks early-stage ventures with a working prototype and a clear path to deployment in health. I have the prototype. I have validation data. I have named partnership discussions with Servier in Suresnes, with Paris-Saclay through the I2BC and Institut Pasteur, and with Sanofi in Gentilly. The accelerator’s structured mentorship and milestone-based progress would help me convert those discussions into a pilot agreement. My roadmap is concrete: fine-tune ESM-2 on the SKEMPI 3K mutation set to reach an AUROC of at least 0.70, then run a pilot with Servier, then build toward recurring revenue. I am applying as a sole founder. That is a risk, and I acknowledge it. The accelerator’s cohort model and mentor network would provide the operational and scientific feedback that a co-founder would otherwise supply. I am already engaged with SEMIA and Quest for Health, and I have submitted applications to IncubAlliance and AI House. Turn.io would be the next step in that pipeline, timed to start after the WILCO One BioTech cohort in October 2026. The venture is pre-seed and not yet incorporated. I am targeting the Ile-de-France region for incorporation and intend to apply for the EIC Accelerator and BPI i-Lab after the Turn.io programme. Non-dilutive grants between 30,000 and 2.5 million euros are the primary funding strategy at this stage. I am asking for a seat in the 2026 cohort. I will bring a validated AI model, a clear health impact thesis in antimicrobial resistance and oncology, and a partnership pipeline that is ready to execute. SHORT ESSAY: INNOVATION AND TECHNICAL FEASIBILITY The innovation is that the model predicts drug resistance mutations from protein sequence alone, without requiring a crystal structure. This is possible because ESM-2, a 650-million-parameter protein language model trained on 250 million sequences, encodes structural and functional information in its internal representations. By computing delta-embeddings between wild-type and mutant sequences, the model captures the effect of a single amino acid substitution on the entire protein landscape. Those delta-embeddings are concatenated with ECFP4 fingerprints of the drug molecule, and a Random Forest classifier makes the final prediction. Technical feasibility is demonstrated by two independent benchmarks. On Platinum, a curated set of 553 mutations across 46 protein-drug complexes, the model achieves AUROC 0.804 with protein-grouped cross-validation. On SKEMPI 2.0, a more challenging set of binding-affinity mutations, the current AUROC is 0.634. The gap between these two numbers defines the immediate technical roadmap: fine-tune ESM-2 on the SKEMPI 3K mutation set, which is three times larger than the current training data, to push SKEMPI performance above 0.70. That target is achievable because the model architecture is proven and the training data bottleneck is the only remaining constraint. The method is computationally lightweight. A single prediction takes under two seconds on a consumer GPU. No cryo-EM, no X-ray crystallography, no molecular dynamics simulations are required. This makes the tool deployable in settings where structural biology infrastructure is limited, including clinical labs in low-resource health systems. SHORT ESSAY: IMPACT AND BUSINESS MODEL The direct health impact is on two disease areas. In oncology, resistance mutations to kinase inhibitors and targeted therapies emerge within months of treatment. A tool that predicts which mutations will arise for a given drug would allow oncologists to select a second-line agent before resistance becomes clinically detectable. In antimicrobial resistance, the same approach applies to antibiotics and antivirals, where resistance is a global health emergency. The business model is a software-as-a-service platform sold to pharmaceutical R&D teams. The primary customer is a computational chemist or biologist in a mid-to-large pharma company who needs to prioritize which mutations to test experimentally. The value proposition is cost reduction: each experimental resistance assay costs between 5,000 and 20,000 euros, and the model can eliminate 60 to 80 percent of those assays by predicting which mutations are unlikely to confer resistance. Revenue model is per-seat annual subscription, with a tier for small biotechs at 15,000 euros per year and an enterprise tier at 50,000 euros per year. The go-to-market strategy is to convert the Servier pilot into a paid contract, then use that reference to approach Sanofi and other Paris-based pharma companies. The total addressable market is the global computational drug discovery market, valued at approximately 4 billion euros and growing at 14 percent annually. Scalability is inherent in the software model. Once the model is fine-tuned and deployed, adding new drugs and new proteins requires only sequence data, not new experiments. The platform can cover any protein-drug pair for which a sequence and a chemical structure are available. CHECKLIST - [ ] Motivation letter, 500 words maximum - [ ] Short essay on innovation and technical feasibility, 350 words maximum - [ ] Short essay on impact and business model, 350 words maximum - [ ] Founder CV, one page - [ ] Proof-of-concept validation data summary, one page - [ ] Partnership confirmation letters from Servier, Paris-Saclay, or Sanofi (if available) - [ ] Incorporation plan for Ile-de-France, one paragraph - [ ] Budget projection for 12 months post-accelerator, one page - [ ] Letter of recommendation from a scientific advisor or previous programme director EDITOR NOTES - Eligibility risk: Turn.io may require a registered company. The venture is not yet incorporated. Confirm whether a pre-incorporation application is accepted, or whether a French micro-enterprise registration would satisfy the requirement. - Fact verification: The AUROC of 0.804 on Platinum is reported as plus or minus 0.025. Confirm the exact standard deviation from the cross-validation runs and whether this is a 95 percent confidence interval or a standard error. - Gap: The founder’s pharmacist background and ML retraining are mentioned but not detailed. The application should include a one-page CV that lists the pharmacy degree, the ML coursework or bootcamp, and any prior startup or research experience. - Partnership status: The profile says discussions are in progress with Servier, Paris-Saclay, and Sanofi. The application should clarify whether any of these have resulted in a signed letter of intent or a data-sharing agreement, or whether they remain informal meetings. - Programme timing: Turn.io 2026 accelerator deadline is May 31, 2026. The WILCO One BioTech cohort is October 2026. Confirm that the founder can participate in both without conflict, or state a preference for Turn.io as the primary programme.